Benchmark Registry

MATH-Vision

MATH-Vision results reported for 9 models across 1 version. Updated 2026-10-03.

Report an issue

Versions

1 version, plus 1 variant or configuration · Metric: Accuracy

VersionReleasedModelsResults
MATH-Vision OriginalLatest
1 variant of Original
VariantConfigurationReleasedModelsResults
MATH-Vision Original — Python enabledWith PythonAugust 6, 202611
February 22, 202499

Recently reported results: MATH-Vision Original

Latest version: MATH-Vision Original

ModelScoreSource
Kimi K3 (max)94.3%raw.githubusercontent.comAccuracyReported Technical reportMoonshot AISelf-reportedRecord
Gemma 4 31B Instruct (thinking mode)85.6%huggingface.coAccuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 12B Instruct (thinking mode)79.7%huggingface.coAccuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 26B A4B Instruct (thinking mode)82.4%huggingface.coAccuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 E4B Instruct (thinking mode)59.5%huggingface.coAccuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord